OpenAI Patches GPT-6 Sol and Luna Vision Bug

OpenAI has deployed a server-side patch to resolve an image encoding defect in its recently released multimodal reasoning models, GPT-6 Sol and GPT-6 Luna. The bug silently degraded visual understanding across the API, Codex, and computer-use workflows. Because the models still completed requests without throwing explicit errors, developers were left with plausible-sounding but inaccurate answers and lower-than-expected evaluation scores.

The defect occurred within the image encoder, which translates pixels into internal representations for model analysis. This flaw caused the models to omit or distort visual details, leading to errors when reading text, identifying objects, or interpreting user interface layouts. The fix is particularly critical for computer-use agents, which rely on precise screenshot analysis to select controls, fill out forms, and track task states.

The update applies to both the high-capability GPT-6 Sol (gpt-6-sol) and the lower-cost, high-volume GPT-6 Luna (gpt-6-luna). For reference, standard API pricing for GPT-6 Sol is $2 per 1 million input tokens and $10 per 1 million output tokens for prompts up to 272,000 input tokens. This is not the first time OpenAI has addressed such an issue; the company deployed a similar fix in March to correct an encoder bug affecting input images in GPT-5.4.

Because the patch was applied directly to OpenAI's servers, developers do not need to make any code changes or migrate model versions. However, since the update alters model behavior, OpenAI recommends that teams rerun vision evaluations conducted between the models' launch and the patch. Developers should also review application-level caches, retry failed visual agent workflows, and re-evaluate any model-selection decisions that were based on earlier, degraded visual performance.